ABSTRACT
This study explores the predictive power of neural networks trained on simple technical indicators, specifically the 5-day moving average (MA5) and exponential moving average (EMA5, when applied to equity portfolios sorted by historical return volatility. Using daily data from the CRSP database spanning from 1927 to 2022, we construct ten volatility deciles and train a feedforward neural network to forecast next-day returns within each decile. Our findings reveal that predictive accuracy and returns are strongest in mid-volatility deciles, particularly decile 7, while performance weakens in the highest-volatility group. These results highlight the importance of risk segmentation in evaluating signal efficacy and suggest that even shallow neural networks using basic technical inputs can outperform traditional strategies when applied selectively. The paper contributes to the literature by introducing a structured, volatility-aware modeling framework that balances simplicity, transparency, and robust predictive performance.
Keywords
Neural networks, artificial intelligence, technical analysis, volatility.